Document final endpoint policy and rejected speaker-tail pilot

README.md

@@ -67,6 +67,21 @@ audio = model.generate(
 sf.write("sample.wav", audio.detach().cpu().numpy(), model.sample_rate)
 ```
 
+## Interactive Demo Endpoint Profile
+
+The hosted [BlueMagpie-TTS Demo](https://huggingface.co/spaces/voidful/BlueMagpie-TTS-Demo)
+uses the same checkpoint weights with an inference-only endpoint guard. It
+starts with a 0.50 stop threshold, lowers the threshold only after 75% of the
+native-rate duration estimate, reaches 0.05 at 95%, and accepts one stop hit.
+Generation is still capped at the native-rate estimate plus one latent step.
+Playback pace is corrected only after generation has stopped, so pace control
+does not hold the model open and create extra speech.
+
+This interactive policy is intentionally separate from the reproducible
+offline evaluation default of `stop_threshold=0.65` and
+`stop_consecutive=2`. Neither policy uses retry selection or best-of-N
+reranking in its standard path.
+
 ## Conditioning Modes
 
 ```python
@@ -161,6 +176,11 @@ did not pass.
 Numbers below are from an internal held-out evaluation set. The eval set and
 training data are intentionally not described in this model card.
 
+The latest residual-speaker pilot improved natural and long-form aggregate
+metrics, but every tested gate scale failed at least one stress endpoint,
+intelligibility, or speaker-boundary criterion. Its weights were not promoted;
+the selected model fingerprint remains unchanged.
+
 | Mode | Evaluation profile | CER | WER |
 | --- | --- | ---: | ---: |
 | centroid-only | 30 texts x 5 seeds, controlled pace | 15.37% | 41.48% |

USAGE.md

@@ -166,6 +166,12 @@ Recommended generation defaults:
 - `continuation-context-sec=0` until generated-context continuation passes its own gate
 - Reference wavs used to extract speaker embeddings should be at least 3 seconds.
 
+The hosted interactive demo uses a separate endpoint-only policy with the same
+weights: a 0.50 stop threshold that relaxes after 75% of the native-rate
+duration estimate to 0.05 at 95%, one stop hit, and a native-rate hard cap plus
+one latent step. It applies pace correction after generation. The offline
+defaults above remain the fixed evaluation contract.
+
 Safety:
 
 - Use only rights-cleared reference audio or speaker embeddings.

release_metadata.json

@@ -33,6 +33,22 @@
         "speaker-reference-embedding"
       ]
     },
+    "interactive_demo": {
+      "scope": [
+        "centroid-only",
+        "speaker-reference-embedding"
+      ],
+      "weight_change": false,
+      "stop_threshold": 0.5,
+      "stop_late_threshold": 0.05,
+      "stop_late_start_ratio": 0.75,
+      "stop_late_full_ratio": 0.95,
+      "stop_consecutive": 1,
+      "hard_stop_margin_steps": 1,
+      "retry_badcase": false,
+      "candidate_rerank": "none",
+      "pace_correction": "post_generation_pitch_preserving"
+    },
     "quality_rerank": {
       "scope": [
         "speaker-reference-embedding-short-medium"